CNN-BiLSTM-Attention Model in Forecasting Wave Height over South-East China Seas

Though numerical wave models have been applied widely to significant height prediction, they consume massive computing memory and their accuracy needs be further improved. In this paper, a two-dimensional (2D) (SWH) prediction model is established for the South East China Seas. The proposed trained by Wave Watch III (WW3) reanalysis data based on convolutional neural network, bi-directional long short-term attention mechanism (CNN-BiLSTM-Attention). It adopts network extract spatial features of original reduce redundant information input into BiLSTM network. Meanwhile, fully associated time series data. Besides, used assign probability weight output layer units, finally, training constructed. Up 24-h experiments are conducted under normal extreme conditions, respectively. Under condition, 3-, 6-, 12- forecasting, mean values correlation coefficients test set 0.996, 0.991, 0.980, 0.945, corresponding root square errors measured at 0.063 m, 0.105 0.172 0.281 typhoon-forced CNN-BiLSTM-Attention typhoon-induced SWH extracted from WW3 For respectively 0.993, 0.983, 0.958, 0.921, averaged 0.159 0.257 0.437 0.555 performs better than that all result suggests algorithm can 2D forecast with higher efficiency.